bipedal_walker

Bipedal walker using SAC with hints

This is the directory containing the code for SAC agent training with or without hints. The hints are provided by another agent, which is initialized by a saved model. This saved model is first created by running SAC, without hints, to solve the BipedalWalker-v3 environment.

The steps to perform are:

• Train an agent for BipedalWalker-v3 as

python main.py --seed 2 --env_name BipedalWalker-v3 --iteration 1500

• Copy the saved models to use later to provide hints

cp learnerq_eval_1_sac_critic.model hinterq_eval_1_sac_critic.model

cp learnerq_eval_2_sac_critic.model hinterq_eval_2_sac_critic.model

cp learnera_eval_sac_actor.model hintera_eval_sac_actor.model

cp learnerreplaymem_sac.model hinterreplaymem_sac.model

• Now you can use hints, for example to solve the BipedalWalkerHardcore-v3 environment

python main.py --seed 4 --env_name BipedalWalkerHardcore-v3 --iteration 3500 --use_hint

You can change the random seeds as you like.

Files provided are:

main.py : main method for SAC

main_td3.py : main method for TD3

net_sac.py : networks, SAC agent, learning using ADMM

net_td3.py : networks, TD3 agent, learning using ADMM

PER.py : prioritized experience replay memory (not used)

display.py : disply the agent in action